DSCSkipNet: An Accuracy-Complexity Trade-off for Effective Potato Disease Identification in Uncontrolled Environments

Potato crop is ranked third in worldwide food consumption, providing a sustenance over a billion people across diverse environments. However, the prevalence of potato pests and diseases significantly diminishes annual yields. With the aim of enabling early diagnosis for timely intervention, this paper focuses on developing an effective method to identify potato leaf diseases under real-world conditions. The proposed approach employs a Convolutional Neural Network (CNN) architecture with an efficient feature extraction mechanism, named DSC-SkipNet, including Depthwise-Separable Convolutions (DSCs), Skip Connections (SCs) and a feature reduction via a Pointwise Convolution (PC). The DSCSkipNet is trained using a recently released dataset containing high-quality images collected under various difficult conditions. The dataset comprises seven classes, encompassing six diseased classes and one healthy class, featuring complex backgrounds, multiple leaves in the image, and irregular acquisition angles. Through meticulous data augmentation, the obtained accuracy exceeds 80% on the testing data using 4-fold cross-validation protocol, all within a lightweight architecture containing only 3.1 million parameters. The achieved performance enables the establishment of a new benchmark and trade-off in terms of accuracy and complexity efficiency, outperforming existing state-of-the-art models.